Towards auditable and intelligent privacy-preserving Record linkage.

Abstract

The Entity Resolution with Privacy Guarantees (REGP) intends to integrate data private/sensitive files from various data sources maintained by different parties. The REGP aims to identify records (for example, people or objects) that represent the same real-world entity in private data sources maintained by different custodians. Due to recent laws and regulations (e.g. General Regulation of Data Protection), PPRL approaches are increasingly required in application areas from the real world, such as health, credit analysis, public policy evaluation and security national. In practical scenarios, the PPRL process needs to deal with efficiency problems, efficiency (call quality) and privacy. For example, the PPRL process needs to run on large data sources (for example, a database containing personal information from government cash distribution and assistance programs), with an accurate classification of entities while protecting the privacy of the information. In this context, this work proposes improvements in the PPRL process with the in order to mitigate some of the bottlenecks of REGP. In particular, this work presents three major contributions to the REGP process: i) a protocol that enables auditability computation performed during REGP, ii) an unsupervised methodology that leverages knowledge of public datasets to train classifier based on in Machine Learning for the REGP, and iii) a new representation, ̧for the PPRL data encoded/anonymized that allow the use of new neuro networks and classifiers of deep learning in the context of PPRL. The present contributions improve several parts of the PPRL process, aiming to make it more easily used in applications of the real world. With the contribution presented in the thesis, we hope to facilitate several applications ̃ from the real world (e.g., medical, epidemiological, and population studies) and reduce legal/bureaucratic efforts to access and process the data, making enforcement of these simpler applications for businesses and governments.